Post #2714711

We've had machine-learned OCR for a good 40 years. It didn't require massively parallel GPUs, huge datacenters, or even what we'd consider powerful machines to train, since they only needed a few hundred samples. A common educational sample I did maybe 20 years ago was training to OCR digits. Any desktop could converge the training pretty quickly from a human comprehendible data set size. I don't remember how many neurons it used.. under 100 IIRC. Not all machine learning is the overly massive stuff we see today. A lot of expert systems were very responsibly trained. It's the push for general that's killing everything.